gsd-pi 2.17.0 → 2.18.0

package/README.md

@@ -21,6 +21,25 @@ One command. Walk away. Come back to a built project with clean git history.
 
 ---
 
+## Documentation
+
+Full documentation is available in the [`docs/`](./docs/) directory:
+
+- **[Getting Started](./docs/getting-started.md)** — install, first run, basic usage
+- **[Auto Mode](./docs/auto-mode.md)** — autonomous execution deep-dive
+- **[Configuration](./docs/configuration.md)** — all preferences, models, git, and hooks
+- **[Token Optimization](./docs/token-optimization.md)** — profiles, context compression, complexity routing (v2.17)
+- **[Cost Management](./docs/cost-management.md)** — budgets, tracking, projections
+- **[Git Strategy](./docs/git-strategy.md)** — worktree isolation, branching, merge behavior
+- **[Working in Teams](./docs/working-in-teams.md)** — unique IDs, shared artifacts
+- **[Skills](./docs/skills.md)** — bundled skills, discovery, custom authoring
+- **[Commands Reference](./docs/commands.md)** — all commands and keyboard shortcuts
+- **[Architecture](./docs/architecture.md)** — system design and dispatch pipeline
+- **[Troubleshooting](./docs/troubleshooting.md)** — common issues, doctor, recovery
+- **[Migration from v1](./docs/migration.md)** — `.planning` → `.gsd` migration
+
+---
+
 ## What Changed From v1
 
 The original GSD was a collection of markdown prompts installed into `~/.claude/commands/`. It relied entirely on the LLM reading those prompts and doing the right thing. That worked surprisingly well — but it had hard limits:
@@ -334,6 +353,26 @@ unique_milestone_ids: true
 | `skill_rules`          | Situational rules for skill routing                                                                   |
 | `unique_milestone_ids` | Uses unique milestone names to avoid clashes when working in teams of people                          |
 
+### Token Optimization (v2.17)
+
+GSD 2.17 introduced a coordinated token optimization system that reduces usage by 40-60% on cost-sensitive workloads. Set a single preference to coordinate model selection, phase skipping, and context compression:
+
+```yaml
+token_profile: budget      # or balanced (default), quality
+```
+
+| Profile | Savings | What It Does |
+|---------|---------|-------------|
+| `budget` | 40-60% | Cheap models, skip research/reassess, minimal context inlining |
+| `balanced` | 10-20% | Default models, skip slice research, standard context |
+| `quality` | 0% | All phases, all context, full model power |
+
+**Complexity-based routing** automatically classifies tasks as simple/standard/complex and routes to appropriate models. Simple docs tasks get Haiku; complex architectural work gets Opus. The classification is heuristic (sub-millisecond, no LLM calls) and learns from outcomes via a persistent routing history.
+
+**Budget pressure** graduates model downgrading as you approach your budget ceiling — 50%, 75%, and 90% thresholds progressively shift work to cheaper tiers.
+
+See the full [Token Optimization Guide](./docs/token-optimization.md) for details.
+
 ### Bundled Tools
 
 GSD ships with 14 extensions, all loaded automatically:

package/dist/onboarding.js

@@ -633,7 +633,7 @@ async function runRemoteQuestionsStep(p, pc, authStorage) {
         });
         if (p.isCancel(channelId) || !channelId)
             return null;
-        const { saveRemoteQuestionsConfig } = await import('./resources/extensions/remote-questions/remote-command.js');
+        const { saveRemoteQuestionsConfig } = await import('./remote-questions-config.js');
         saveRemoteQuestionsConfig('slack', channelId.trim());
         p.log.success(`Slack channel: ${pc.green(channelId.trim())}`);
         return 'Slack';
@@ -736,7 +736,7 @@ async function runDiscordChannelStep(p, pc, token) {
         channelId = channelChoice;
     }
     // Save remote questions config
-    const { saveRemoteQuestionsConfig } = await import('./resources/extensions/remote-questions/remote-command.js');
+    const { saveRemoteQuestionsConfig } = await import('./remote-questions-config.js');
     saveRemoteQuestionsConfig('discord', channelId);
     const channelName = channels.find(ch => ch.id === channelId)?.name;
     p.log.success(`Discord channel: ${pc.green(channelName ? `#${channelName}` : channelId)}`);

package/dist/remote-questions-config.d.ts

@@ -0,0 +1,10 @@
+/**
+ * Remote Questions Config Helper
+ *
+ * Extracted from remote-questions extension so onboarding.ts can import
+ * it without crossing the compiled/uncompiled boundary. The extension
+ * files in src/resources/ are shipped as raw .ts and loaded via jiti,
+ * but onboarding.ts is compiled by tsc — dynamic imports from compiled
+ * JS to uncompiled .ts fail at runtime (#592).
+ */
+export declare function saveRemoteQuestionsConfig(channel: "slack" | "discord", channelId: string): void;

package/dist/remote-questions-config.js

@@ -0,0 +1,36 @@
+/**
+ * Remote Questions Config Helper
+ *
+ * Extracted from remote-questions extension so onboarding.ts can import
+ * it without crossing the compiled/uncompiled boundary. The extension
+ * files in src/resources/ are shipped as raw .ts and loaded via jiti,
+ * but onboarding.ts is compiled by tsc — dynamic imports from compiled
+ * JS to uncompiled .ts fail at runtime (#592).
+ */
+import { existsSync, readFileSync, writeFileSync, mkdirSync } from "node:fs";
+import { dirname } from "node:path";
+import { getGlobalGSDPreferencesPath } from "./resources/extensions/gsd/preferences.js";
+export function saveRemoteQuestionsConfig(channel, channelId) {
+    const prefsPath = getGlobalGSDPreferencesPath();
+    const block = [
+        "remote_questions:",
+        `  channel: ${channel}`,
+        `  channel_id: "${channelId}"`,
+        "  timeout_minutes: 5",
+        "  poll_interval_seconds: 5",
+    ].join("\n");
+    const content = existsSync(prefsPath) ? readFileSync(prefsPath, "utf-8") : "";
+    const fmMatch = content.match(/^---\n([\s\S]*?)\n---/);
+    let next = content;
+    if (fmMatch) {
+        let frontmatter = fmMatch[1];
+        const regex = /remote_questions:[\s\S]*?(?=\n[a-zA-Z_]|\n---|$)/;
+        frontmatter = regex.test(frontmatter) ? frontmatter.replace(regex, block) : `${frontmatter.trimEnd()}\n${block}`;
+        next = `---\n${frontmatter}\n---${content.slice(fmMatch[0].length)}`;
+    }
+    else {
+        next = `---\n${block}\n---\n\n${content}`;
+    }
+    mkdirSync(dirname(prefsPath), { recursive: true });
+    writeFileSync(prefsPath, next, "utf-8");
+}

package/dist/resources/extensions/gsd/activity-log.ts

@@ -8,7 +8,7 @@
  * Diagnostic extraction is handled by session-forensics.ts.
  */
 
-import { writeFileSync, mkdirSync, readdirSync, unlinkSync, statSync, openSync, closeSync, constants } from "node:fs";
+import { writeFileSync, writeSync, mkdirSync, readdirSync, unlinkSync, statSync, openSync, closeSync, constants } from "node:fs";
 import { createHash } from "node:crypto";
 import { join } from "node:path";
 
@@ -23,6 +23,15 @@ interface ActivityLogState {
 
 const activityLogState = new Map<string, ActivityLogState>();
 
+/**
+ * Clear accumulated activity log state (#611).
+ * Call when auto-mode stops to prevent unbounded memory growth
+ * from lastSnapshotKeyByUnit maps accumulating across units.
+ */
+export function clearActivityLogState(): void {
+  activityLogState.clear();
+}
+
 function scanNextSequence(activityDir: string): number {
   let maxSeq = 0;
   try {
@@ -46,9 +55,21 @@ function getActivityState(activityDir: string): ActivityLogState {
   return state;
 }
 
-function snapshotKey(unitType: string, unitId: string, content: string): string {
-  const digest = createHash("sha1").update(content).digest("hex");
-  return `${unitType}\0${unitId}\0${digest}`;
+/**
+ * Build a lightweight dedup key from session entries without serializing
+ * the entire content to a string (#611). Uses entry count + hash of
+ * the last few entries as a fingerprint instead of hashing megabytes.
+ */
+function snapshotKey(unitType: string, unitId: string, entries: unknown[]): string {
+  const hash = createHash("sha1");
+  hash.update(`${unitType}\0${unitId}\0${entries.length}\0`);
+  // Hash only the last 3 entries as a fingerprint — if the session grew,
+  // the count change alone detects it; if content changed, the tail hash catches it.
+  const tail = entries.slice(-3);
+  for (const entry of tail) {
+    hash.update(JSON.stringify(entry));
+  }
+  return hash.digest("hex");
 }
 
 function nextActivityFilePath(
@@ -91,14 +112,23 @@ export function saveActivityLog(
     mkdirSync(activityDir, { recursive: true });
 
     const safeUnitId = unitId.replace(/\//g, "-");
-    const content = `${entries.map(entry => JSON.stringify(entry)).join("\n")}\n`;
     const state = getActivityState(activityDir);
     const unitKey = `${unitType}\0${safeUnitId}`;
-    const key = snapshotKey(unitType, safeUnitId, content);
+    // Use lightweight fingerprint instead of serializing all entries (#611)
+    const key = snapshotKey(unitType, safeUnitId, entries);
     if (state.lastSnapshotKeyByUnit.get(unitKey) === key) return;
 
     const filePath = nextActivityFilePath(activityDir, state, unitType, safeUnitId);
-    writeFileSync(filePath, content, "utf-8");
+    // Stream entries to disk line-by-line instead of building one massive string (#611).
+    // For large sessions, the single-string approach allocated hundreds of MB.
+    const fd = openSync(filePath, "w");
+    try {
+      for (const entry of entries) {
+        writeSync(fd, JSON.stringify(entry) + "\n");
+      }
+    } finally {
+      closeSync(fd);
+    }
     state.nextSeq += 1;
     state.lastSnapshotKeyByUnit.set(unitKey, key);
   } catch (e) {

package/dist/resources/extensions/gsd/auto-prompts.ts

@@ -89,7 +89,7 @@ export async function inlineDependencySummaries(
 export async function inlineGsdRootFile(
   base: string, filename: string, label: string,
 ): Promise<string | null> {
-  const key = filename.replace(/\.md$/i, "").toUpperCase() as "PROJECT" | "DECISIONS" | "QUEUE" | "STATE" | "REQUIREMENTS";
+  const key = filename.replace(/\.md$/i, "").toUpperCase() as "PROJECT" | "DECISIONS" | "QUEUE" | "STATE" | "REQUIREMENTS" | "KNOWLEDGE";
   const absPath = resolveGsdRootFile(base, key);
   if (!existsSync(absPath)) return null;
   return inlineFileOptional(absPath, relGsdRootFile(key), label);
@@ -377,6 +377,8 @@ export async function buildResearchMilestonePrompt(mid: string, midTitle: string
   if (requirementsInline) inlined.push(requirementsInline);
   const decisionsInline = await inlineGsdRootFile(base, "decisions.md", "Decisions");
   if (decisionsInline) inlined.push(decisionsInline);
+  const knowledgeInlineRM = await inlineGsdRootFile(base, "knowledge.md", "Project Knowledge");
+  if (knowledgeInlineRM) inlined.push(knowledgeInlineRM);
   inlined.push(inlineTemplate("research", "Research"));
 
   const inlinedContext = `## Inlined Context (preloaded — do not re-read these files)\n\n${inlined.join("\n\n---\n\n")}`;
@@ -413,6 +415,8 @@ export async function buildPlanMilestonePrompt(mid: string, midTitle: string, ba
   if (requirementsInline) inlined.push(requirementsInline);
   const decisionsInline = inlineLevel !== "minimal" ? await inlineGsdRootFile(base, "decisions.md", "Decisions") : null;
   if (decisionsInline) inlined.push(decisionsInline);
+  const knowledgeInlinePM = await inlineGsdRootFile(base, "knowledge.md", "Project Knowledge");
+  if (knowledgeInlinePM) inlined.push(knowledgeInlinePM);
   inlined.push(inlineTemplate("roadmap", "Roadmap"));
   if (inlineLevel === "full") {
     inlined.push(inlineTemplate("decisions", "Decisions"));
@@ -461,6 +465,8 @@ export async function buildResearchSlicePrompt(
   if (decisionsInline) inlined.push(decisionsInline);
   const requirementsInline = await inlineGsdRootFile(base, "requirements.md", "Requirements");
   if (requirementsInline) inlined.push(requirementsInline);
+  const knowledgeInlineRS = await inlineGsdRootFile(base, "knowledge.md", "Project Knowledge");
+  if (knowledgeInlineRS) inlined.push(knowledgeInlineRS);
   inlined.push(inlineTemplate("research", "Research"));
 
   const depContent = await inlineDependencySummaries(mid, sid, base);
@@ -504,6 +510,8 @@ export async function buildPlanSlicePrompt(
     const requirementsInline = await inlineGsdRootFile(base, "requirements.md", "Requirements");
     if (requirementsInline) inlined.push(requirementsInline);
   }
+  const knowledgeInlinePS = await inlineGsdRootFile(base, "knowledge.md", "Project Knowledge");
+  if (knowledgeInlinePS) inlined.push(knowledgeInlinePS);
   inlined.push(inlineTemplate("plan", "Slice Plan"));
   if (inlineLevel === "full") {
     inlined.push(inlineTemplate("task-plan", "Task Plan"));
@@ -578,11 +586,16 @@ export async function buildExecuteTaskPrompt(
     ? priorSummaries.slice(-1)
     : priorSummaries;
   const carryForwardSection = await buildCarryForwardSection(effectivePriorSummaries, base);
+
+  // Inline project knowledge if available
+  const knowledgeInlineET = await inlineGsdRootFile(base, "knowledge.md", "Project Knowledge");
+
   const inlinedTemplates = inlineLevel === "minimal"
     ? inlineTemplate("task-summary", "Task Summary")
     : [
         inlineTemplate("task-summary", "Task Summary"),
         inlineTemplate("decisions", "Decisions"),
+        ...(knowledgeInlineET ? [knowledgeInlineET] : []),
       ].join("\n\n---\n\n");
 
   const taskSummaryPath = `${relSlicePath(base, mid, sid)}/tasks/${tid}-SUMMARY.md`;
@@ -624,6 +637,8 @@ export async function buildCompleteSlicePrompt(
     const requirementsInline = await inlineGsdRootFile(base, "requirements.md", "Requirements");
     if (requirementsInline) inlined.push(requirementsInline);
   }
+  const knowledgeInlineCS = await inlineGsdRootFile(base, "knowledge.md", "Project Knowledge");
+  if (knowledgeInlineCS) inlined.push(knowledgeInlineCS);
 
   // Inline all task summaries for this slice
   const tDir = resolveTasksDir(base, mid, sid);
@@ -697,6 +712,8 @@ export async function buildCompleteMilestonePrompt(
     const projectInline = await inlineGsdRootFile(base, "project.md", "Project");
     if (projectInline) inlined.push(projectInline);
   }
+  const knowledgeInlineCM = await inlineGsdRootFile(base, "knowledge.md", "Project Knowledge");
+  if (knowledgeInlineCM) inlined.push(knowledgeInlineCM);
   // Inline milestone context file (milestone-level, not GSD root)
   const contextPath = resolveMilestoneFile(base, mid, "CONTEXT");
   const contextRel = relMilestoneFile(base, mid, "CONTEXT");
@@ -825,6 +842,8 @@ export async function buildReassessRoadmapPrompt(
     const decisionsInline = await inlineGsdRootFile(base, "decisions.md", "Decisions");
     if (decisionsInline) inlined.push(decisionsInline);
   }
+  const knowledgeInlineRA = await inlineGsdRootFile(base, "knowledge.md", "Project Knowledge");
+  if (knowledgeInlineRA) inlined.push(knowledgeInlineRA);
 
   const inlinedContext = `## Inlined Context (preloaded — do not re-read these files)\n\n${inlined.join("\n\n---\n\n")}`;
 

package/dist/resources/extensions/gsd/auto-worktree.ts

@@ -14,8 +14,10 @@ import {
   removeWorktree,
   worktreePath,
 } from "./worktree-manager.js";
+import { detectWorktreeName } from "./worktree.js";
 import {
   MergeConflictError,
+  readIntegrationBranch,
 } from "./git-service.js";
 import { parseRoadmap } from "./files.js";
 import { loadEffectiveGSDPreferences } from "./preferences.js";
@@ -90,7 +92,12 @@ export function autoWorktreeBranch(milestoneId: string): string {
  */
 export function createAutoWorktree(basePath: string, milestoneId: string): string {
   const branch = autoWorktreeBranch(milestoneId);
-  const info = createWorktree(basePath, milestoneId, { branch });
+
+  // Use the integration branch recorded in META.json as the start point.
+  // This ensures the worktree branch is created from the branch the user
+  // was on when they started the milestone (e.g. f-setup-gsd-2), not main.
+  const integrationBranch = readIntegrationBranch(basePath, milestoneId) ?? undefined;
+  const info = createWorktree(basePath, milestoneId, { branch, startPoint: integrationBranch });
 
   // Copy .gsd/ planning artifacts from the source repo into the new worktree.
   // Worktrees are fresh git checkouts — untracked files don't carry over.
@@ -224,6 +231,27 @@ export function getAutoWorktreeOriginalBase(): string | null {
   return originalBase;
 }
 
+export function getActiveAutoWorktreeContext(): {
+  originalBase: string;
+  worktreeName: string;
+  branch: string;
+} | null {
+  if (!originalBase) return null;
+  const cwd = process.cwd();
+  const resolvedBase = existsSync(originalBase) ? realpathSync(originalBase) : originalBase;
+  const wtDir = join(resolvedBase, ".gsd", "worktrees");
+  if (!cwd.startsWith(wtDir)) return null;
+  const worktreeName = detectWorktreeName(cwd);
+  if (!worktreeName) return null;
+  const branch = nativeGetCurrentBranch(cwd);
+  if (!branch.startsWith("milestone/")) return null;
+  return {
+    originalBase,
+    worktreeName,
+    branch,
+  };
+}
+
 // ─── Merge Milestone -> Main ───────────────────────────────────────────────
 
 /**
@@ -279,11 +307,12 @@ export function mergeMilestoneToMain(
   const previousCwd = process.cwd();
   process.chdir(originalBasePath_);
 
-  // 4. Resolve main branch from preferences
+  // 4. Resolve integration branch — prefer milestone metadata, fall back to preferences / "main"
   const prefs = loadEffectiveGSDPreferences()?.preferences?.git ?? {};
-  const mainBranch = prefs.main_branch || "main";
+  const integrationBranch = readIntegrationBranch(originalBasePath_, milestoneId);
+  const mainBranch = integrationBranch ?? prefs.main_branch ?? "main";
 
-  // 5. Checkout main
+  // 5. Checkout integration branch
   nativeCheckoutBranch(originalBasePath_, mainBranch);
 
   // 6. Build rich commit message

package/dist/resources/extensions/gsd/auto.ts

@@ -29,7 +29,7 @@ import {
   buildMilestoneFileName, buildSliceFileName, buildTaskFileName,
 } from "./paths.js";
 import { invalidateAllCaches } from "./cache.js";
-import { saveActivityLog } from "./activity-log.js";
+import { saveActivityLog, clearActivityLogState } from "./activity-log.js";
 import { synthesizeCrashRecovery, getDeepDiagnostic } from "./session-forensics.js";
 import { writeLock, clearLock, readCrashLock, formatCrashInfo, isLockProcessAlive } from "./crash-recovery.js";
 import {
@@ -92,7 +92,9 @@ import {
   getAutoWorktreePath,
   getAutoWorktreeOriginalBase,
   mergeMilestoneToMain,
+  autoWorktreeBranch,
 } from "./auto-worktree.js";
+import { pruneQueueOrder } from "./queue-order.js";
 import { showNextAction } from "../shared/next-action-ui.js";
 import {
   resolveExpectedArtifactPath,
@@ -196,6 +198,33 @@ function shouldUseWorktreeIsolation(): boolean {
   return true; // default: worktree
 }
 
+/**
+ * Detect and escape a stale worktree cwd (#608).
+ *
+ * After milestone completion + merge, the worktree directory is removed but
+ * the process cwd may still point inside `.gsd/worktrees/<MID>/`.
+ * When a new session starts, `process.cwd()` is passed as `base` to startAuto
+ * and all subsequent writes land in the wrong directory. This function detects
+ * that scenario and chdir back to the project root.
+ *
+ * Returns the corrected base path.
+ */
+function escapeStaleWorktree(base: string): string {
+  const marker = `${pathSep}.gsd${pathSep}worktrees${pathSep}`;
+  const idx = base.indexOf(marker);
+  if (idx === -1) return base;
+
+  // base is inside .gsd/worktrees/<something> — extract the project root
+  const projectRoot = base.slice(0, idx);
+  try {
+    process.chdir(projectRoot);
+  } catch {
+    // If chdir fails, return the original — caller will handle errors downstream
+    return base;
+  }
+  return projectRoot;
+}
+
 /** Crash recovery prompt — set by startAuto, consumed by first dispatchNextUnit */
 let pendingCrashRecovery: string | null = null;
 
@@ -228,6 +257,9 @@ const DISPATCH_GAP_TIMEOUT_MS = 5_000; // 5 seconds
 /** SIGTERM handler registered while auto-mode is active — cleared on stop/pause. */
 let _sigtermHandler: (() => void) | null = null;
 
+/** Tool calls currently being executed — prevents false idle detection during long-running tools. */
+const inFlightTools = new Set<string>();
+
 type BudgetAlertLevel = 0 | 75 | 90 | 100;
 
 export function getBudgetAlertLevel(budgetPct: number): BudgetAlertLevel {
@@ -293,6 +325,22 @@ export function isAutoPaused(): boolean {
   return paused;
 }
 
+/**
+ * Mark a tool execution as in-flight. Called from index.ts on tool_execution_start.
+ * Prevents the idle watchdog from declaring the agent idle while tools are executing.
+ */
+export function markToolStart(toolCallId: string): void {
+  if (!active) return;
+  inFlightTools.add(toolCallId);
+}
+
+/**
+ * Mark a tool execution as completed. Called from index.ts on tool_execution_end.
+ */
+export function markToolEnd(toolCallId: string): void {
+  inFlightTools.delete(toolCallId);
+}
+
 /**
  * Return the base path to use for the auto.lock file.
  * Always uses the original project root (not the worktree) so that
@@ -345,6 +393,7 @@ function clearUnitTimeout(): void {
     clearInterval(idleWatchdogHandle);
     idleWatchdogHandle = null;
   }
+  inFlightTools.clear();
   clearDispatchGapWatchdog();
 }
 
@@ -426,14 +475,18 @@ export async function stopAuto(ctx?: ExtensionContext, pi?: ExtensionAPI): Promi
         `Auto-worktree teardown failed: ${err instanceof Error ? err.message : String(err)}`,
         "warning",
       );
-      // Force basePath back to original even if teardown failed
-      if (originalBasePath) {
-        basePath = originalBasePath;
-        try { process.chdir(basePath); } catch { /* best-effort */ }
-      }
     }
   }
 
+  // Always restore cwd to project root on stop (#608).
+  // Even if isInAutoWorktree returned false (e.g., module state was already
+  // cleared by mergeMilestoneToMain), the process cwd may still be inside
+  // the worktree directory. Force it back to originalBasePath.
+  if (originalBasePath) {
+    basePath = originalBasePath;
+    try { process.chdir(basePath); } catch { /* best-effort */ }
+  }
+
   const ledger = getLedger();
   if (ledger && ledger.units.length > 0) {
     const totals = getProjectTotals(ledger.units);
@@ -458,12 +511,15 @@ export async function stopAuto(ctx?: ExtensionContext, pi?: ExtensionAPI): Promi
   stepMode = false;
   unitDispatchCount.clear();
   unitRecoveryCount.clear();
+  inFlightTools.clear();
   lastBudgetAlertLevel = 0;
   unitLifetimeDispatches.clear();
   currentUnit = null;
   currentMilestoneId = null;
   originalBasePath = "";
+  completedUnits = [];
   clearSliceProgressCache();
+  clearActivityLogState();
   pendingCrashRecovery = null;
   _handlingAgentEnd = false;
   ctx?.ui.setStatus("gsd-auto", undefined);
@@ -519,6 +575,11 @@ export async function startAuto(
 ): Promise<void> {
   const requestedStepMode = options?.step ?? false;
 
+  // Escape stale worktree cwd from a previous milestone (#608).
+  // After milestone merge + worktree removal, the process cwd may still point
+  // inside .gsd/worktrees/<MID>/ — detect and chdir back to project root.
+  base = escapeStaleWorktree(base);
+
   // If resuming from paused state, just re-activate and dispatch next unit.
   // The conversation is still intact — no need to reinitialize everything.
   if (paused) {
@@ -569,17 +630,17 @@ export async function startAuto(
     ctx.ui.setFooter(hideFooter);
     ctx.ui.notify(stepMode ? "Step-mode resumed." : "Auto-mode resumed.", "info");
     // Restore hook state from disk in case session was interrupted
-    restoreHookState(base);
+    restoreHookState(basePath);
     // Rebuild disk state before resuming — user interaction during pause may have changed files
-    try { await rebuildState(base); } catch { /* non-fatal */ }
+    try { await rebuildState(basePath); } catch { /* non-fatal */ }
     try {
-      const report = await runGSDDoctor(base, { fix: true });
+      const report = await runGSDDoctor(basePath, { fix: true });
       if (report.fixesApplied.length > 0) {
         ctx.ui.notify(`Resume: applied ${report.fixesApplied.length} fix(es) to state.`, "info");
       }
     } catch { /* non-fatal */ }
     // Self-heal: clear stale runtime records where artifacts already exist
-    await selfHealRuntimeRecords(base, ctx, completedKeySet);
+    await selfHealRuntimeRecords(basePath, ctx, completedKeySet);
     invalidateAllCaches();
     await dispatchNextUnit(ctx, pi);
     return;
@@ -1251,6 +1312,11 @@ async function dispatchNextUnit(
     unitLifetimeDispatches.clear();
     // Capture integration branch for the new milestone and update git service
     captureIntegrationBranch(originalBasePath || basePath, mid, { commitDocs: loadEffectiveGSDPreferences()?.preferences?.git?.commit_docs });
+    // Prune completed milestone from queue order file
+    const pendingIds = state.registry
+      .filter(m => m.status !== "complete")
+      .map(m => m.id);
+    pruneQueueOrder(basePath, pendingIds);
   }
   if (mid) {
     currentMilestoneId = mid;
@@ -1331,6 +1397,39 @@ async function dispatchNextUnit(
           `Milestone merge failed: ${err instanceof Error ? err.message : String(err)}`,
           "warning",
         );
+        // Ensure cwd is restored even if merge failed partway through (#608).
+        // mergeMilestoneToMain may have chdir'd but then thrown, leaving us
+        // in an indeterminate location.
+        if (originalBasePath) {
+          basePath = originalBasePath;
+          try { process.chdir(basePath); } catch { /* best-effort */ }
+        }
+      }
+    } else if (currentMilestoneId && !isInAutoWorktree(basePath)) {
+      // Branch isolation mode (#603): no worktree, but we may be on a milestone/* branch.
+      // Squash-merge back to the integration branch (or main) before stopping.
+      try {
+        const currentBranch = getCurrentBranch(basePath);
+        const milestoneBranch = autoWorktreeBranch(currentMilestoneId);
+        if (currentBranch === milestoneBranch) {
+          const roadmapPath = resolveMilestoneFile(basePath, currentMilestoneId, "ROADMAP");
+          if (roadmapPath) {
+            const roadmapContent = readFileSync(roadmapPath, "utf-8");
+            // mergeMilestoneToMain handles: auto-commit, checkout integration branch,
+            // squash merge, commit, optional push, branch deletion.
+            const mergeResult = mergeMilestoneToMain(basePath, currentMilestoneId, roadmapContent);
+            gitService = new GitServiceImpl(basePath, loadEffectiveGSDPreferences()?.preferences?.git ?? {});
+            ctx.ui.notify(
+              `Milestone ${currentMilestoneId} merged (branch mode).${mergeResult.pushed ? " Pushed to remote." : ""}`,
+              "info",
+            );
+          }
+        }
+      } catch (err) {
+        ctx.ui.notify(
+          `Milestone merge failed (branch mode): ${err instanceof Error ? err.message : String(err)}`,
+          "warning",
+        );
       }
     }
     sendDesktopNotification("GSD", `Milestone ${mid} complete!`, "success", "milestone");
@@ -1757,6 +1856,10 @@ async function dispatchNextUnit(
         startedAt: currentUnit.startedAt,
         finishedAt: Date.now(),
       });
+      // Cap to last 200 entries to prevent unbounded growth (#611)
+      if (completedUnits.length > 200) {
+        completedUnits = completedUnits.slice(-200);
+      }
       clearUnitRuntimeRecord(basePath, currentUnit.type, currentUnit.id);
       unitDispatchCount.delete(`${currentUnit.type}/${currentUnit.id}`);
       unitRecoveryCount.delete(`${currentUnit.type}/${currentUnit.id}`);
@@ -1957,6 +2060,16 @@ async function dispatchNextUnit(
     if (!runtime) return;
     if (Date.now() - runtime.lastProgressAt < idleTimeoutMs) return;
 
+    // Agent has tool calls currently executing (await_job, long bash, etc.) —
+    // not idle, just waiting for tool completion.
+    if (inFlightTools.size > 0) {
+      writeUnitRuntimeRecord(basePath, unitType, unitId, currentUnit.startedAt, {
+        lastProgressAt: Date.now(),
+        lastProgressKind: "tool-in-flight",
+      });
+      return;
+    }
+
     // Before triggering recovery, check if the agent is actually producing
     // work on disk.  `git status --porcelain` is cheap and catches any
     // staged/unstaged/untracked changes the agent made since lastProgressAt.